Merge pull request #12 from PaddlePaddle/develop

00

docs/appendix/model_zoo.md

@@ -6,28 +6,28 @@
 
 | 模型  | 模型大小 | 预测速度（毫秒） | Top1准确率（%） | Top5准确率（%） |
 | :----|  :------- | :----------- | :--------- | :--------- |
-| ResNet18| 46.9MB   | 1.499        | 71.0     | 89.9     |
-| ResNet34| 87.5MB   | 2.272        | 74.6    | 92.1    |
-| ResNet50| 102.7MB  | 2.939        | 76.5     | 93.0     |
-| ResNet101 |179.1MB  | 5.314      | 77.6     | 93.6  |
-| ResNet50_vd |102.8MB  | 3.165        | 79.1     | 94.4     |
-| ResNet101_vd| 179.2MB  | 5.252       | 80.2   | 95.0     |
-| ResNet50_vd_ssld |102.8MB  | 3.165        | 82.4     | 96.1     |
-| ResNet101_vd_ssld| 179.2MB  | 5.252       | 83.7   | 96.7     |
-| DarkNet53|166.9MB  | 3.139       | 78.0     | 94.1     |
-| MobileNetV1 | 16.0MB   | 32.523        | 71.0     | 89.7    |
-| MobileNetV2 | 14.0MB   | 23.318        | 72.2     | 90.7    |
-| MobileNetV3_large|  21.0MB   | 19.308        | 75.3    | 93.2   |
-| MobileNetV3_small |  12.0MB   | 6.546        | 68.2    | 88.1     |
-| MobileNetV3_large_ssld|  21.0MB   | 19.308        | 79.0     | 94.5     |
-| MobileNetV3_small_ssld |  12.0MB   | 6.546        | 71.3     | 90.1     |
-| Xception41 |92.4MB   | 4.408       | 79.6    | 94.4     |
-| Xception65 | 144.6MB  | 6.464       | 80.3     | 94.5     |
-| DenseNet121 | 32.8MB   | 4.371       | 75.7     | 92.6     |
-| DenseNet161|116.3MB  | 8.863       | 78.6     | 94.1     |
-| DenseNet201|  84.6MB   | 8.173       | 77.6     | 93.7     |
-| ShuffleNetV2 | 9.0MB   | 10.941        | 68.8     | 88.5     |
-| HRNet_W18 | 21.29MB | 7.368 (V100 GPU) | 76.9 | 93.4 |
+| ResNet18| 46.9MB   | -        | 71.0     | 89.9     |
+| ResNet34| 87.5MB   | -        | 74.6    | 92.1    |
+| ResNet50| 102.7MB  | -        | 76.5     | 93.0     |
+| ResNet101 |179.1MB  | -      | 77.6     | 93.6  |
+| ResNet50_vd |102.8MB  |-        | 79.1     | 94.4     |
+| ResNet101_vd| 179.2MB  | -       | 80.2   | 95.0     |
+| ResNet50_vd_ssld |102.8MB  | -        | 82.4     | 96.1     |
+| ResNet101_vd_ssld| 179.2MB  | -       | 83.7   | 96.7     |
+| DarkNet53|166.9MB  | -       | 78.0     | 94.1     |
+| MobileNetV1 | 16.0MB   | -        | 71.0     | 89.7    |
+| MobileNetV2 | 14.0MB   | -        | 72.2     | 90.7    |
+| MobileNetV3_large|  21.0MB   | -        | 75.3    | 93.2   |
+| MobileNetV3_small |  12.0MB   | -        | 68.2    | 88.1     |
+| MobileNetV3_large_ssld|  21.0MB   | -        | 79.0     | 94.5     |
+| MobileNetV3_small_ssld |  12.0MB   | -        | 71.3     | 90.1     |
+| Xception41 |92.4MB   | -       | 79.6    | 94.4     |
+| Xception65 | 144.6MB  | -       | 80.3     | 94.5     |
+| DenseNet121 | 32.8MB   | -       | 75.7     | 92.6     |
+| DenseNet161|116.3MB  | -       | 78.6     | 94.1     |
+| DenseNet201|  84.6MB   | -       | 77.6     | 93.7     |
+| ShuffleNetV2 | 9.0MB   | -        | 68.8     | 88.5     |
+| HRNet_W18 | 21.29MB | - | 76.9 | 93.4 |
 
 ## 目标检测模型
 

paddlex/cv/models/hrnet.py

@@ -77,6 +77,7 @@ class HRNet(DeepLabv3p):
         self.class_weight = class_weight
         self.ignore_index = ignore_index
         self.labels = None
+        self.fixed_input_shape = None
 
     def build_net(self, mode='train'):
         model = paddlex.cv.nets.segmentation.HRNet(
@@ -86,7 +87,8 @@ class HRNet(DeepLabv3p):
             use_bce_loss=self.use_bce_loss,
             use_dice_loss=self.use_dice_loss,
             class_weight=self.class_weight,
-            ignore_index=self.ignore_index)
+            ignore_index=self.ignore_index,
+            fixed_input_shape=self.fixed_input_shape)
         inputs = model.generate_inputs()
         model_out = model.build_net(inputs)
         outputs = OrderedDict()
@@ -170,6 +172,6 @@ class HRNet(DeepLabv3p):
         return super(HRNet, self).train(
             num_epochs, train_dataset, train_batch_size, eval_dataset,
             save_interval_epochs, log_interval_steps, save_dir,
-            pretrain_weights, optimizer, learning_rate, lr_decay_power,
-            use_vdl, sensitivities_file, eval_metric_loss, early_stop,
+            pretrain_weights, optimizer, learning_rate, lr_decay_power, use_vdl,
+            sensitivities_file, eval_metric_loss, early_stop,
             early_stop_patience, resume_checkpoint)

paddlex/cv/models/load_model.py

@@ -116,20 +116,7 @@ def fix_input_shape(info, fixed_input_shape=None):
         resize = {'ResizeByShort': {}}
         padding = {'Padding': {}}
         if info['_Attributes']['model_type'] == 'classifier':
-            crop_size = 0
-            for transform in info['Transforms']:
-                if 'CenterCrop' in transform:
-                    crop_size = transform['CenterCrop']['crop_size']
-                    break
-            assert crop_size == fixed_input_shape[
-                0], "fixed_input_shape must == CenterCrop:crop_size:{}".format(
-                    crop_size)
-            assert crop_size == fixed_input_shape[
-                1], "fixed_input_shape must == CenterCrop:crop_size:{}".format(
-                    crop_size)
-            if crop_size == 0:
-                logging.warning(
-                    "fixed_input_shape must == input shape when trainning")
+            pass
         else:
             resize['ResizeByShort']['short_size'] = min(fixed_input_shape)
             resize['ResizeByShort']['max_size'] = max(fixed_input_shape)

paddlex/cv/models/slim/prune_config.py

@@ -142,13 +142,16 @@ def get_prune_params(model):
     program = model.test_prog
     if model_type.startswith('ResNet') or \
             model_type.startswith('DenseNet') or \
-            model_type.startswith('DarkNet'):
+            model_type.startswith('DarkNet') or \
+            model_type.startswith('AlexNet'):
         for block in program.blocks:
             for param in block.all_parameters():
                 pd_var = fluid.global_scope().find_var(param.name)
                 pd_param = pd_var.get_tensor()
                 if len(np.array(pd_param).shape) == 4:
                     prune_names.append(param.name)
+        if model_type == 'AlexNet':
+            prune_names.remove('conv5_weights')
     elif model_type == "MobileNetV1":
         prune_names.append("conv1_weights")
         for param in program.global_block().all_parameters():

paddlex/cv/nets/segmentation/hrnet.py

@@ -38,7 +38,8 @@ class HRNet(object):
                  use_bce_loss=False,
                  use_dice_loss=False,
                  class_weight=None,
-                 ignore_index=255):
+                 ignore_index=255,
+                 fixed_input_shape=None):
         # dice_loss或bce_loss只适用两类分割中
         if num_classes > 2 and (use_bce_loss or use_dice_loss):
             raise ValueError(
@@ -66,6 +67,7 @@ class HRNet(object):
         self.use_dice_loss = use_dice_loss
         self.class_weight = class_weight
         self.ignore_index = ignore_index
+        self.fixed_input_shape = fixed_input_shape
         self.backbone = paddlex.cv.nets.hrnet.HRNet(
             width=width, feature_maps="stage4")
 
@@ -131,8 +133,16 @@ class HRNet(object):
 
     def generate_inputs(self):
         inputs = OrderedDict()
-        inputs['image'] = fluid.data(
-            dtype='float32', shape=[None, 3, None, None], name='image')
+
+        if self.fixed_input_shape is not None:
+            input_shape = [
+                None, 3, self.fixed_input_shape[1], self.fixed_input_shape[0]
+            ]
+            inputs['image'] = fluid.data(
+                dtype='float32', shape=input_shape, name='image')
+        else:
+            inputs['image'] = fluid.data(
+                dtype='float32', shape=[None, 3, None, None], name='image')
         if self.mode == 'train':
             inputs['label'] = fluid.data(
                 dtype='int32', shape=[None, 1, None, None], name='label')